Networking off Madison Avenue1 Mohammad Arzaghi and J. Vernon Henderson

Brown University

October 12, 2006

New York County, or Manhattan, accounts for 24% of advertising agency receipts in the

USA. Years ago the popular image had large agencies on Madison Avenue generating this

business. But today there are over 1000 agencies in different clusters spread over the southern

half of Manhattan. In clustering, advertising is known for the key role that networking plays in

the operation of agencies, where interviews suggest that informal networking with exchange of

ideas and expertise among close neighbors is critical to many agencies’ success (Arzaghi, 2005).

We interpret localized networking as a structured case of information spillovers, or Marshall’s

(1890) “mysteries” of the “air”, involving more deliberate information exchanges.

This paper examines the effect on productivity of having more near advertising agency

neighbors and hence better opportunities for meetings and information exchange within

Manhattan. We have census tract data for Manhattan that allows us to distinguish locations at

250 meter increments. With this fine level of geographic detail, we can infer the high benefits of

close social interactions. We show that there is extremely rapid spatial decay in the benefits of

more near neighbors even in the close quarters of southern Manhattan, a finding that is new to

the empirical literature. This suggests that having a high density of similar commercial

establishments is important in enhancing local productivity for certain industries, as modeled in

Lucas and Rossi-Hansberg (2002).

Our results indicate that, in Manhattan, advertising agencies trade-off the higher rent

costs of being in bigger clusters closer to the “centers of action”, against the lower rent costs of

operating on the “fringes” away from high concentrations of other agencies. That trade-off

indicates that the benefits of higher spillovers, or better opportunities to network, within

metropolitan areas such as Manhattan, may be largely capitalized into rents, rather than wages.

1 The research in this paper was conducted while the authors were Census Bureau research associates at the Boston Research Data Center. Research results and conclusions expressed are those of the authors and do not necessarily indicate concurrence by the Census Bureau. The paper has been screened to ensure that no confidential data are revealed. We acknowledge helpful comments from Andrew Foster and William Strange on aspects of the paper and from participants in seminars at Brown University, LSE, NBER Real Estate and Local Public Finance Meetings, Yale University, University of Minnesota and University of Toronto.

2

This study differs in key ways from and challenges aspects of the traditional work on

urban scale externalities. First, almost the entire literature, as reviewed in Rosenthal and Strange

(2004), examines externalities in manufacturing. Manufacturing to the extent it still plays an

important role in American city productivity is found disproportionately in rural areas and

smaller and medium size cities (Kolko, 1999). If we want to understand the raison d’être of

larger metropolitan areas and why certain industries are drawn to high rent, high wage cities like

New York with historically built dense neighborhoods, we need to look at high-end business and

financial services, such as advertising services.

Second, the literature on externalities examines either total factor productivity benefits

(e.g., Henderson, 2003) or wage gains (e.g., Glaeser and Mare, 2001, and Rosenthal and Strange,

2005) of greater agglomeration at the aggregate county or metropolitan area level. One problem

just noted is that localized information spillovers may be capitalized into commercial rents, not

wages. Second key aspects of information spillovers are captured by within county, not across

county variation. For advertising agencies, networking benefits are so localized, that they are

exhausted within the confines of sub-areas of southern Manhattan: firms on the far west side of

southern Manhattan derive no direct benefits from firms on the east side. While Rosenthal and

Strange (2003) introduce spatial decay into externality estimation, they look just at

manufacturing (and software) and assume no spatial decay within the first mile, while here

effects completely decay within a half mile.

Third, this literature debates what the nature of scale benefits is—information spillovers,

labor market externalities, or reduced transactions and transport costs of exchange of

intermediate inputs as in the new economic geography (Fujita, Krugman, and Venables, 2001)—

with no clear resolution. Looking within a county allows us to isolate specific sources of

externalities. We know what we capture are not labor market externalities: Southern Manhattan

is all one labor market, drawing workers from all over the New York PMSA and even beyond.2

We will argue below that finding immediate spatial proximity to other advertising agencies

matters implies proximity benefits for information exchange and networking, rather than some

2 There is evidence of very modest wage gradients (over longer distances than just within Manhattan) in large metro areas. But it is ascribed to the fact that those living further from the center earn lower wages because they work locally and have lower commuting costs (supply side) and that clusters further from the city center may have lower scale benefits (demand side), rather than there being different labor markets within a metro area (McMillan and Singell, 1992).

3

type of shopping-center externality (Dudey, 1990) or market potential effect. We also look for

address signaling effects, such as a particular benefit from locating in the Madison Avenue area.

If what we isolate within Manhattan involves information spillovers, that still leaves the

question of how to interpret cross county or metropolitan area results which find aggregate scale

of own industry activity matters. For advertising agencies, they could represent labor market or

shopping-center externalities at the metropolitan area level; but they might also represent greater

opportunities for clustering and networking in larger metropolitan areas. A research challenge for

future work is how to sort out the spatial scales at which different types of agglomeration

benefits operate.

In terms of scale externalities outside the own industry, we explore the benefits of

immediate access to buyers (e.g., headquarters), suppliers (e.g., graphics) and outlets

(broadcasters). Only access to broadcasters matters, which represents the new economic

geography costs of trade (negotiating and buying advertising time). But immediate proximity to

headquarters or graphic services does not matter. The latter are ubiquitous and the former are a

short cab ride away in Manhattan where day-to-day, face-to-face interaction is not required. In

fact, as discussed below, most advertising services produced in Manhattan are now exported to

companies away from New York City.

Finally, a potential advantage of looking within counties is that cross MSA and county

studies are plagued by unmeasured characteristics which affect both productivity and local

industry scale and vary across large geographic units. These include legal, regulatory and

business climate, access to specific markets, and specific infrastructure including the transport

system which affects commuting and work effort. In a panel context one can try to control for

these with fixed effects (Henderson, 2003), but such effects may not be time invariant, especially

over longer panels. Southern Manhattan is all one legal-regulatory-business climate and

infrastructure system, so the usual cross-section unobservables are not an issue in identification.

There are tract level unobservables; but we argue that by their nature they are easier to deal with

econometrically, so that identification is more straightforward than in MSA level analysis.

1 Advertising Agencies In this section we give some background information on the advertising agency industry,

in general and in Manhattan. Then we look at Manhattan in more detail.

4

1.1 General Background

In 1997, advertising expenditures in the U.S. totaled $187b, about 2.2% of GDP. Of this,

only $17b involves advertising agency receipts (NIPA, BEA).3 However of the $17b, $10b

comes from commissions on $89b of media billings carried out by advertising agencies. This

implies that over half of advertising expenditures in the U.S. go through advertising agencies and

reflects the fact that many advertising agencies continue the historical function of intermediation,

or matchmaking between advertisers and media outlets. However, advertising agencies are more

popularly known for the design of advertising campaigns. From the industry point of view, top

notch firms provide “great creative thinking campaigns and ideas” for advertisers to market their

products and to develop a “unique selling proposition” (Schemetter, 2003).

Advertising agencies operate with internal teams, where a team typically provides the full

range of services for a client—market research, creation of an advertising campaign, production,

media, client management, and the like, where most agency costs are labor costs. Apart from

potential media commissions, advertising agencies derive a significant portion of their income

from fees and billings for creative services rendered in the design of campaigns.

To win an account, an agency team competes by developing new ideas. Thus, like R&D

(Jaffe et al., 1993), with creativity playing a central role, information sharing and information

diffusion are thought to be critical to an agency’s success. That sharing occurs within and across

agencies, where agencies belong to formal and informal networks.4 Based on interviews with

agencies in New York and Chicago, informal networking appears to occur at the design stage of

advertising campaigns where an agency has received a “request for proposal” (or RFP) for an

advertising campaign (Arzaghi, 2004). The agency turns to members of its local network for

information (who by odds are unlikely to have received the same RFP) on how best to respond to

the request for proposal and how to design the potential campaign, as well as turning to its

network to borrow materials and research. In New York, we believe clustering occurs because of

3 Advertising is split about 55-45 between national versus local advertising (Berndt and Silk, 1994) and is focused on automobiles (18%), retail, department and discount stores (15%) and then movies, cosmetics and toiletries, medicines, food, financial services and restaurants with shares each of 4.3-5%. In terms of media, TV, radio and cable TV account for 36%, magazines and newspapers 32% and yellow pages and direct mail 30%. The last two sets of numbers are from the AdAge website, based on reports by respectively TNS Media Intelligence and Coen/McCann-Erikson. Both data sets are property of Crain Communication, Inc.. 4 Data exist on the formal network structure of advertising agencies, where members pay dues to networks which have full-time managers and they attend network meetings several times a year. Hameroff (1998) lists eleven huge national and international formal networks of larger agencies. These formal networks primarily exchange information about marketing conditions across cities and countries.

5

localized networking to enhance creativity; agencies share information, ideas, and materials

where repeated face-to-face contact is critical. Contacts may be “official” such as meetings

among members of a neighborhood network, unofficial such as informal contacts in local lunch

and coffee places, or both.

The notion that networking is so important to advertising agencies compared to most

other activities may be related to the fact that advertising agencies have low industrial

concentration, in terms of both establishment and firm size. Table 1 compares advertising

agencies relative with business services more generally, noting business services are much less

concentrated than manufacturing. Establishments under five employees account for 58% of all

establishments and 14% of sales, higher than for business services generally. The average

establishment size is 11 employees compared to about 50 in manufacturing and to 21 in the rest

of business services. Single-unit firms account for 90% of establishments and 55% of sales in the

advertising agency industry, compared to 81% and 45% in other business services.

While the advertising agency industry is a non-concentrated industry, it is heavily

spatially concentrated in New York County, or Manhattan, as shown in Table 2. In 1997,

Manhattan has 7.3% of advertising agency establishments in the U.S., 20% of their employment,

24% of all advertising agency receipts, and 31% of media billings. For comparison, we note that

Manhattan has about 1.9% of all private employment in the USA; so location quotients for New

York for advertising agency activities are very high. The relatively fewer numbers of

establishments (7.3%) compared to employment (20%) means establishment sizes are much

larger in Manhattan than elsewhere. The high concentration of advertising agencies in Manhattan

means these agencies “export” much of their product; their buyers are outside Manhattan (noting

also that Manhattan’s share of headquarters has declined significantly over time, now garnering

only 3% of national headquarters employment).

In the empirical work, in examining the advertising agency business, we focus on the

activities of single-unit [SU] firms, as opposed to establishments of multi-unit [MU] firms.

Multi-unit firms are huge firms with multiple, very large establishments, which seem to serve

different types of clients. In New York, SU’s have an average size of 11 employees, while MU

establishments (not firms) have an average size of 155 employees. MU’s are much more reliant

on income from media receipts indicating their greater role as intermediaries between advertisers

and the media. In 1997 in NY, they got 48% of receipts from media billings, while SU’s got

6

28%. In contrast, SU’s are more involved in the creative design of products, with much more

income from fees for services. They also seem to engage much more in networking. In 1997,

they got 21% of their income on commissions and fees for services provided by other agencies,

compared to just 6% for MU’s, an indication of greater networking, or “sharing” through cross-

firm indirect sales by SU’s. Examples of such cross-agency services cited by the Census

questionnaire include artwork, plates, and market and other research.

Why don’t we combine the groups, and investigate MU’s more thoroughly? First in the

text and Appendix B, we argue that, in econometric work, firms in each group seem to interact

within the own group (SU and MU), but not across groups. In the type of Pearson 2χ tests

discussed later, it is also the case that MU and SU establishment stocks have different location

patterns across zip codes in all years of our data. This lack of interaction presumably relates to

our inference that they operate in different markets. Second, sample sizes for MU’s are 10% of

those for SU’s. In pooling, SU effects dominate all results (i.e., we get basically the same results

as in the Tables below) and making detailed inferences separately for MU’s is difficult (and

paper lengths are limited). Finally, we don’t expect cross firm networking to be so important for

MU’s; for their more media based activities, MU’s rely on intra-firm networks across their own

establishments.

Before proceeding we note there are a whole set of industrial organization questions we

do not have the information in Census data to deal with. We look at location patterns of births of

new firms to infer what location attributes, including networking potential they value. There is a

dynamic of new firms being split-offs where an employee leaves an existing firm to start a new

firm and then a later dynamic of deaths of firms which include buy-outs. We don’t have any

information on this process. Nor do we examine (or observe) how multi-unit firms arise through

acquisitions versus creation of new establishments; nor the segmentation of the market between

single and multi-unit firms. We are just focused on inferring networking benefits of single unit

firms, to document the nature and speed of spatial decay of such benefits. For that purpose, the

main issue is heterogeneity of benefits to different types of firms (to the extent births have

information on an underlying productivity parameter that might affect their net-working benefits

(see Section 2.1)). We have no ex ante information on births. However in the last section of the

paper, we explore heterogeneity issues using a small sample of movers.

7

1.2 Manhattan

Within Manhattan, advertising agencies are closely clustered. Because of disclosure

criteria, we are not allowed to map agency locations using Census data. Although the empirical

work to follow is based on the entire population of advertising agencies in the Census Business

Register (or Standard Statistical Establishment List—SSEL), we use Advertising Red Books

online for 2003 to generate maps. The Red Books give the location of about 800 of the 1000 or

so agencies in Manhattan.5 Figure 1 shows advertising agency locations in the southern half of

Manhattan, defined as the area below 90th Street. Almost all agencies in NY are located below

90th Street. Overall, clustering is heaviest within a block or two of Madison and Fifth Avenues,

but there appear to be many “sub-clusters” as we move north-south along the Avenues, looking

both east and west. The Red Books do not reliably breakdown single and multi-unit firms, so we

cannot map them separately. If we compare Figure 1 with the rent gradient map in Figure A-2,

we can see there are large clusters of agencies in medium rent areas, with somewhat fewer in the

highest rent areas. In low rent areas on the fringes, there are even fewer agencies. For 1992, the

simple correlation between the number of single unit advertising agencies in a tract and the log

of rent price is .23.

In order to further explore the spatial distribution of agencies within Manhattan, we

return to Census data. We examine similarity of location patterns by zip code, asking whether

any comparison pair (e.g., stocks in 1977 versus stocks in 1992) is drawn from the same

distribution (across zip codes). We use zip code location information here, rather than the Census

tracts we use later, since only the former are available from the Census in all years (1977, 82, 87,

92, 97). Using Pearson 2χ tests6 to compare location patterns across zip codes for all of

Manhattan, we examine a key aspect of location patterns, relevant in the sections to follow.

Location patterns of single unit advertising agencies in Manhattan have changed overtime, which

is useful information in discussing identification later. Specifically:

(1) Comparing 1987 (or 1982) with 1997, the stock location pattern of single unit

firms in 1987 differs from the pattern in 1997. 5 This is a publication of National Register Publishing. The Red Book includes self reported data.

6 The test statistic is 2

1 1

( )

r cij ij

i j ij

m eT

e= =

−= ∑∑ where ji

ij i j

nne Np p NN N

⋅⋅= ≡ is the ML estimate of probability of cell

ij where in ⋅ is sum across rows (say, and )SU MU and jn⋅ is sum across columns (say, locations), and ijm is the

count in any cell. 2 is distributed T χ with degree of freedom (r-1) (c-1).

8

(2) Corresponding to (1), births in 1982-87 vs. 1992-97 have different spatial

distributions.7

Patterns change from 1987 to 1997 reputedly in response to relative rental cost changes

across areas of Manhattan at that time, moving advertising agencies out of some of the highest

rent areas. Such relative rent cost changes must be correlated with and probably driven by

changing tract characteristics over time, an issue we will return to below. Most births of single

unit firms since 1992 have happened in medium and lower rent areas and hence off Madison

Avenue.

2. A Model of Expected Networking and Location Choice In this section, we develop a model of location choice within New York City for single-

unit advertising agencies, to help us think about empirical issues. We start with the specification

of networking technology and location choice. In section 3, we turn to the details of

implementation, empirical specification, data, estimation, and results. Section 4 provides

conclusions and analyzes some extensions.

2.1 Formulating Local Network Contacts

We specify a simple model for a new firm that is choosing a location in anticipation of

informally networking with other agencies nearby. The distance between agencies plays a key

role in the extent of networking. While the specification and jargon is informed by the large

theoretical literature on networks (e.g., Jackson and Wolinsky, 1996, Bala and Goyal, 2000, and

Goyal and Moraga-Gonzalez, 2002), we look at just one firm and how to specify the benefits,

costs and extent of networking for it. We model networking as overlapping with direct linkages

only, where each pair of establishments decides on its pair-wise extent of contact, independent of

other pair-wise arrangements.

The idea is that a new firm anticipates networking with advertising agencies in close-by

locations, where this networking can involve social “meetings” with employees of other agencies

(e.g., going to the bar after work or coffee in the morning), scheduled business meetings where

some team members in a firm get together with some team members in another firm, or contacts

7 These differences apply to establishments of multi-unit firms as well; but as we already above their location patterns differ from those of single unit firms.

9

through intermediaries such as local graphics firms or contract designers serving several

agencies. A new firm in networking with any other firm has a match parameter, revealed after

the firm has picked its location and established contact with the other firm. Given the ex post

chosen level of networking for each pair, networking is costly in terms of time, where that cost

rises with distance between pairs. This could represent commuting time for the requisite face-to-

face meetings or monitoring of one another’s activities.

Although our estimation is not strictly structural, we provide a framework to motivate our

specification and interpret results for those who want a more structural interpretation. Consider a

firm operating at location j . The value of its contact with a firm at location i is

[ c( )]j ij ij ij ijA V u V Fγ δ− − . (1)

In (1), ijV is the volume of communications j gets from a firm at i and 0 1,γ< < so there are

decreasing returns to volume in any pair of contacts. iju is a “match” parameter drawn on the

interval [0,1], which is firm specific between the firm at j and a firm at i, and gives variation in

the extent of contacts across firms. ( )ijc δ is the cost per unit of communication, where ijδ is the

distance from to i j . Total communication costs comprise firm labor costs; one can define ijV as

labor units involved in communication, with ( )ijc δ representing lost work time from employees

having to travel further to interact with another firm. Differences in jA could represent

differences across locations in amenities. F is the initial fixed cost of making a contact with a

firm at another location, upon which iju is learned.8

In formulating (1) and in interpreting results, as noted above, advertising agencies mostly

sell to firms outside Manhattan and initial “shopping” of advertisers for advertising agencies is

done by mail, telephone, and internet, not by “walk-in-traffic”. Thus, unlike a retail concern, we

don’t expect market potential effects or shopping-center externalities within Manhattan.

Nonetheless, in the empirical formulation we will examine and discuss formulations that allow

for shopping-center externalities and/or market potential effects.

8 While iju can take the value zero to represent “no match” possibility, if we want to incorporate issues of conflict of interest across agencies, we would want to have a probability p of any contact resulting in “no match” and formulate an expected fixed cost per contact as ( )1F p− , with only actual retained contacts contributing to variable profits.

10

In our formulation, once the firm spends F to learn iju , it chooses ijV so that

* 1/1 1/1ij ( /c( )) ( )ij ijV uγ γγ δ − −= . If ( ) ( )ij jic cδ δ= and ij jiu u= , then in this simple case ij jiV V= and

communications are reciprocal and two sided. In the present specification,

ex post profit from contact with any firm is /(1 ) (1 ) 1/(1 )(1 ) ( ) ( )j ij ijA c u Fγ γ γ γ γγ γ δ− − − −− − and ex ante

profit is /(1 ) (1 ) 1/(1 )(1 ) ( ) [ ]j ijA c E u Fγ γ γ γ γγ γ δ− − − −− − (2),

where we denote the expected value of 1/(1 )( )iju γ− as 1/(1 )[ ]E u γ− , which we assume is common

across all i , for now. Note the expected volumes of communications and profits are decreasing

in distance and communication costs.

Equation (2) relates to the key set of coefficients we will attempt to estimate—how

expected profit to a firm at j is increased if a firm is added at i and how that profit declines with

distance, ijδ . Equation (2) also can be used to define the maximum distance, maxδ , over which a

firm makes contacts and scale externalities are relevant. Empirically we will see that there is

such a cut-off point. Given the positive term in (2) is declining in ijδ , maxδ is defined by

/(1 ) (1 ) 1/(1 )max max{ | (1 ) ( ) [ ] 0}ij j ijA c E u Fγ γ γ γ γδ δ γ γ δ− − − −= − − > . (3)

maxδ is increasing in firm productivity, jA , and in expected match productivity 1/(1 )[ ]E u γ− , and is

declining in costs.

In this simple context, total firm profits are simply the value of all contacts the firm

makes minus the costs. Thus total expected profits are

max1

1 1 1maxE[ ] (1 ) E[ ] ( ) ( )j j ij j

iA u c FN R

γ γδγ γ γγ γ δ δ

−− − −Π = − − −∑ , (4)

where jR is the rental cost at location j and ( )maxδN is the number of agencies within the

distance maxδ of location j . Given the definition of maxδ in (3), it is clear that having more firms

within maxδ (i.e., greater the density) raises profits. This presents two issues: why is there more

than one cluster in New York and how might we start to think about heterogeneity of firms?

Multiple clusters. Why are there multiple clusters in Manhattan? First, in empirical

specifications, the strategy is to have a firm-location specific matching parameter drawn in the

usual logit fashion, so firms have preferences over locations per se, and thus spread out, some in

11

distinct clusters and some more in isolation. Second, in Manhattan the number of available

spaces within a neighborhood is limited and it isn’t clear there could be a neighborhood of 1000

agencies. Third, one could specify the model so that there are decreasing returns to scale to a

firm in overall volume of all communications limiting the extent of contacts; Arzaghi (2004)

models optimal network sizes as advertising agency compete for accounts, even without spatial

decay. Finally, another motivation for multiple clusters is heterogeneity, the empirics of we

discuss in the last section.

Heterogeneity. A context where there is variation in what different quality firms are willing to

pay for features of advertising agency clusters has the potential to deliver a separating

equilibrium as in Arzaghi (2005). There, high quality firms cluster in high cost neighborhoods

and lower quality ones operate on the fringes because they can’t afford the higher costs in larger

clusters. Differences in firm quality could be reflected by differences across firms inγ or in the

specification of matching parameters. For the latter we could assume each firm is endowed with

its own “communication ability” parameter ju , a measure of firm quality. For any pair, the

match gets min[ , ], where ij j i ju u u u= is the communication parameter for our firm at j and iu is

the parameter for a specific firm at i . Given [0,1]u∈ , one can show that

1/(1 )1/(1 ) /(1 )

0

1[( ) ] ( ) ( )1

ju

jE u u u G u duγγ γ γ

γ−− −= −

−∫ , (5)

where ( )G u is the distribution function. Equation (5) suggests higher quality firms have

higher 1/(1 )[( ) ]E u γ− ; and they benefit more from neighbors in equations (2) or (4), so that their

scale effects are enhanced. They also benefit more from being in a neighborhood with better

quality neighbors, in the sense of a better distribution function, ( )G u .9 Higher quality firms will

have more profits and be larger (more volume of communications) and thus in our interpretation

will have larger employment. In the last section of the paper when we introduce firm

heterogeneity, we will use size to measure firm quality.

9 If a firm chooses a neighborhood with a better distribution, ( )F u , of neighbors in the sense of first degree

stochastic dominance, then 1/(1 )[( ) ]E u γ− rises by /(1 )

0

1 ( ( ) ( )) 01

ju

u G u F u duγ γ

γ− − >

−∫ . This rise in 1/(1 )[( ) ]E u γ− is

greater for higher quality firms by 1 /(1 )(1 ) ( ) ( ( ) ( ))j j ju G u F uγ γγ − −− − .

12

2.2 Empirical Implementation and Data

The model we estimate is based on equation (4). In the treatment of space, locations are

census tracts, which in NY are at a finer level of spatial detail than zip codes. Moreover, a census

tract is intended to represent a true micro “neighborhood,” as most tracts in Manhattan coincide

with the city blocks between two avenues and a few streets. Each census tract has an interpolated

rental cost per square foot of class A office space, availability of commercial sites measured by

the stock of private establishments, and distance measures to other advertising agencies and other

activities such as broadcasting establishments (where ads are placed), headquarters (clients), or

graphic services (inputs).

In the empirical implementation of equation (4) we want to allow office space sizes to

respond to rental price, rather than being fixed. We adjust our specification of production

technology, to yield a simple estimating equation. If jl is the amount of space consumed at a unit

rental cost, jR , suppose ex post profits are specified as

max

exp {[ c( )] }j j ij ij ij ij ji

A l V u V F R lδ

α γ δ⎛ ⎞

− − −⎜ ⎟⎝ ⎠∑ .

Optimizing with respect to choice of jV gives

max/(1 ) (1 ) 1/(1 )exp {(1 ) ( ) }j j ij ij j j

iA l c u F R l

δα γ γ γ γ γγ γ δ− − − −⎛ ⎞

− − −⎜ ⎟⎝ ⎠∑ .

For the choice of jl , we take expected values, maximize ex ante profits, and substitute back in

for the optimal value of jl . Moreover, given the discrete nature of our geographical units

(tracts), we allow there to be ni firms at each location (tract) and then take the log of expected

profits. After rearrangement, expected profits from choosing location j are

max1

1 1 1j

1ln ln ln E[exp{(1 ) ( ) }] 1 1j j j i ij

iC R n c u F

γ γδγ γ γαπ γ γ δ

α α

−− − −≡ Π = − + − −

− − ∑ . (6)

The empirics focus on estimating the effect of an increase in the number of firms at

location i on the profits of a firm located at j, or a set of coefficients that decline with distance

ijδ , and match the expression /1 /1 1/1ln E[exp{(1 ) ( ) ]ijc u Fγ γ γ γ γγ γ δ− − − −− − in (6). Our estimation is

not structural: to make it structural, we would need to assume a specific PDF for iju and shape of

( )ijc δ , for which we do not have good priors. Moreover, the model is meant to motivate the

13

empirical formulation and interpretation, rather than insist on an exact structure. Finally in (6),

jC contains jA which in turn includes a firm-tract specific error term.

Formulation. In estimation we look at the location choices of births, or new firms in

Manhattan, where firms pick the census tract which maximizes profits. In viewing location

choices across tracts, as usual, we start by assuming that the firm looks at each tract as “one”

choice with a tract-firm idiosyncratic match parameter, jη , drawn from an extreme value type I

distribution. Tracts are defined to be neighborhoods in New York. In a logit formulation the

probability that a firm chooses tract j is

( )( )∑

=

ii

jjP

ππ

expexp

. (7)

In (7), jπ is jπ in (6) without the jC which includes the separable error term, jη 10.

In all specifications, we augment the arguments of jπ in (6) by adding the log of the total

number of private establishments in the neighborhood, im , for two reasons. First, it captures

localized general agglomeration economies from having more commercial activities in a firm’s

own neighborhood, including any general market potential or shopping-center externalities

beyond the specifics of headquarter locations in Manhattan which are treated separately.

However Holmes (2005) suggests a second consideration. Within the own tract, unobserved to

us, spaces differ and rather than assuming the same error draw for a firm within a tract, we could

assume each available location, z , in tract j is associated with its own idiosyncratic matching

parameter to the firm, jzε , drawn from a generalized extreme value distribution. Office spaces on

the same city-block in New York may be somewhat similar (in the same building) and certainly

face similar street conditions, so it is natural to assume that these error drawings are correlated

within tracts. A feature of Manhattan is that tracts differ in available spaces, where land and

building usage is tightly regulated and relatively inflexible over time (Glaeser and Gyourko,

2002). If (unobserved) vacancies are proportional to existing commercial spaces and there are

more available commercial spaces in a tract, then a firm enjoys more idiosyncratic draws and

better matching possibilities within a tract. In that case, the coefficient of jm is 1 σ− , where σ

10 That is, ( )ln j

jC Ceη= .

14

is approximately the within tract correlation of error drawings in a nested logit framework.

Under this interpretation, equation (7) represents the upper level nesting, with the effects of

lower level choices summarized in the effect of jm (given that we do not observe any

characteristics of lower level matches).

With homogeneous births, all covariates are tract characteristics and are not firm specific.

In this case, if the expected number of births in a neighborhood, jλ , takes the usual form

( )[ ]jj mln1exp σπ −+ , the problem may be equivalently estimated as a standard Poisson model

of birth counts per tract (Guimaraes, Figueiredo, and Woodward, 2000). We start by presenting

ordinary Poisson results, but that leaves the problem that error draws for neighborhoods may be

correlated with some neighborhood covariates. As such, we want instruments that influence these

covariates but are exogenous to contemporaneous error draws. Assuming such instruments exist,

we can estimate a (non-linear) moment condition. This also relaxes the Poisson assumption on

mean and variance and takes care of possible over- or under-dispersion problems.

If we define

( )[ ]jjjjjj mBB ln1exp σπλυ −+−=−≡ ,

where is thejB actual number of births and jλ the expected number, the moment condition is

| 0j jE Zυ⎡ ⎤ =⎣ ⎦ , (8)

where jZ are instruments. We use non-linear generalized method of moments suggested in

Windmeijer and Silva (1997) and Mullahy (1997) to estimate the parameters. Of course, if jZ

are simply the covariates of the model, estimation of (8) gives identical coefficients to those in

the Poisson and logit.

Data. We use Census Business Register [SSEL] which covers all US establishments. We

analyze the location decisions of births in Manhattan from 1992 to 1997. A birth is a new firm to

Manhattan—a firm that did not exist in the 1992 SSEL but did in the 1997 one. We chose the

SSEL in Census years because in those years additional information from the Census is available

for firms; and more critically the records on existence and location are updated and most

accurate for Economic Census years. We examine births in the 164 census tracts south of 90th

Street, which cover 97.4% of all single unit advertising agencies. We exclude the 132 tracts north

15

of 90th Street because we cannot accurately infer rent data per tract for them. Correspondingly,

many of those tracts are not viable locations for advertising agencies.

Total births in Manhattan are given in Table 3. For our tracts south of 90th, the 1992 stock

is 949 and the number of births is 502. We base 1992-1997 births on 1992 covariates. This

specification implies that these births have naive expectations concerning economic magnitudes,

which generally fits the data best. However, there is a lot of turnover in firms and their locations

so firms born in, say, the last year (1997) may be basing their decisions on much more updated

information than 1992. Thus based on 1997 Census data which contain a year of birth, we also

look at a sample of births for the 148 establishments who in 1997 give their birth year as 1993 or

1994. [Yearly SSEL records for 1993 and 1994 show almost no births, reflecting the fact that

birth records are not updated until Census years.] We first show that results for the two samples

are the same and then focus on 1992-97 births because the tract birth numbers are more

numerous and hence less noisy.

In terms of covariates, rents are based on zip code asking rents for class A office space in

Manhattan in 1992; we utilize the rents in zip codes for which we have data and use a GIS rent-

contour fitting routine to infer rents for all census tracts in southern Manhattan (see Appendix

A). Tract rental prices vary threefold within southern Manhattan.

In terms of ad agency neighbors, for each tract, we construct measures of access of that

tract to nearby advertising agencies by allocating neighbor agencies in 1992 relative to that tract

into rings. We define 5 rings moving out in increments of 250 meters. Ring 1 is the count of

existing single unit advertising agencies in census tracts whose centroid is within 250 meters of

the own census tract centroid. Mostly ring 1 is just the own census tract (but it can include up to

3 tracts). Ring 2 is the count of agencies in census tracts whose centroids are between 250 and

500 meters of the own tract, and so on for ring’s 3-5, up to 1250 meters. Each additional ring

typically contains 3-4 more tracts. We also experiment with other ring divisions and report one

set of results.11 Finally, we calculate the total number of commercial establishments for each

tract and ring in Manhattan, as well as a variety of other specific commercial activities of

relevance to the advertising business.

11 Note southern Manhattan is about 2600 meters across and about 11,000 meters long.

16

Identification. By focusing just on Manhattan, we minimize the problem in cross- MSA

estimation that localities have different unobserved characteristics in terms of labor markets,

transport infrastructure, and regulatory-business-political climates. Now, the problem reduces to

one of unobserved neighborhood characteristics within Manhattan, such as what the trendy

eating and socializing spots for networking are, where construction is underway making face-to-

face meetings more time costly, or where pick-pocketing, street people and the like are currently

a more troublesome problem for employees and perspective clients. For these types of

unobservable attributes, we believe it is realistic to assume they are constantly changing, so

unobservable, relevant attributes in 1992 are uncorrelated with those 15 years earlier. As such,

we choose historical variables from 1977 as instruments for 1992 covariates. We will report

many tests concerning this assumption, but granting us that for now, the issue is what is

endogenous and what are possible instruments.

In the empirical work, the coefficient subject to the greatest bias will turn out to be rents.

The idea is that the effects of current neighborhood unobservables on businesses in general are

capitalized into rents. Better conditions are reflected in higher rents, biasing the rent coefficient

upwards. Given Manhattan’s highly regulated housing market, commercial rent has strong

instruments: total housing units in 1970, and share of those units in 1970 in buildings with 5 or

fewer units. These variables reflect the potential, despite regulations, to convert residences to

office spaces. In first stage regressions for rent, these variables have significant negative

coefficients. We also control for distance to Rockefeller center as the commercial center of the

city (which itself has no direct effects on advertising agency profits), but its effect in first stage

regressions is weak.

The stocks of advertising agencies in rings are also potentially endogenous variables.

Trendy local conditions not only influence rents and births today, but also may have enough

persistence into the past to affect some of the accumulation of current stocks. The primary

concern is with the own ring stock, since the error drawing is for the own tract in the own ring.

Of course error drawings across tracts and rings may be correlated. While we will test for cross-

tract spatial correlation and show that is not a compelling issue, we still instrument for all ring

17

stock variables. For instruments we use 1977 values of ring stocks.12 Use of these instruments

raises inter-related issues.

If tract unobservables and location patterns of advertising agency stocks do not change

over time, use of historical stocks to instrument for current stocks is not valid. Location patterns

replicate over time with births replacing deaths by tract and with instruments being correlated

with current unobservables. We have already documented that location patterns did change from

the 1980’s to the 1990’s. The driving force cited is relative changes in rent patterns, but such

changes imply underlying changes in tract characteristics. But if that is correct then why are

historical stocks strong instruments for current stocks? The basic argument is that, while there

are distinct shifts in location patterns, within historical clusters there is also some degree of

persistence in location of older firms, driven by moving costs. These include physical moving

costs, loss of address recognition, and the costs of leaving (defecting from) an existing network

(co-ordination failure of moving clusters en masse). In general 1977 ring stocks are strong

instruments, although surprisingly less so for the own ring.

Despite this argument, readers should remain suspicious, a priori. Besides formal and

informal specification tests, we conduct many experiments. These include (a) adding various

measures of other tract characteristics that may affect location choices, (b) looking at patterns of

deaths to show birth patterns are not replicating death patterns, and (c) trying an alternative

instrument for the own ring stock.

For the last, an alternative instrument is based on a dramatic change in the way

headquarters and advertising agencies interact. Not only have headquarters moved away from

Manhattan over the last 20 years, but 20 years ago advertising agencies in Manhattan in their

location choices clung to headquarter locations. In 1977 the simple correlation coefficient

between numbers of advertising agencies and headquarters per tract was .82. By 1992, the

corresponding correlation coefficient had fallen to .28.13 With the internet, cheap air travel,

cheap telephone service and the like, the advertising agency-headquarter interaction has changed.

Now, day-to-day interactions are not face-to-face; and specifically in the highly intense design

stage of advertising campaigns, agencies usually fly a team to headquarter locations away from

12 Actually we can only assign about 70% of establishments in 1977 to a 1992 census tract; and it is the counts of these 70% that we use. 13 Note patterns of residential versus commercial locations across Manhattan generate positive correlation between almost any two business activities

18

Manhattan for some weeks. Given the historical need to be near headquarters and some

persistence in locations of older advertising agencies, the correlation coefficient between 1992

single unit advertising agency location patterns and 1977 headquarter location patterns is .56,

double the coefficient with contemporaneous headquarter location patterns. This makes 1977

headquarter counts by tract a potentially good alternative instrument for 1992 stocks, especially

if one believes headquarters’ location choices in 1977 were little influenced by where advertising

agencies located, being more dependent on tract characteristics not relevant to advertising

agencies (like access to financial services).

3. Results In this section we start with results and identification issues for the basic birth

specification in Table 4. Then we turn to Table 5 which pertains to identification as well. Table 6

explores effects of tract characteristics not included in the basic specification, and Tables 7a and

7b examine robustness. Finally, we consider the problem of spatial correlation.

3.1 Basic Results

Table 4 contains our main results. The discussion is divided into three parts. First we

discuss the effects of IV versus ordinary Poisson estimations, then we turn to the numerical

interpretation of coefficients, and finally we discuss in great detail the validity of our

instruments. The first two columns of Table 4 are for births in 1993 and 1994 (as recorded in the

1997 SSEL), which should be most sensitive to 1992 conditions; and the next two columns are

for 1992-97 births. For 1993 and 1994 births, column 1 contains ordinary Poisson results

(identical to ordinary conditional logit results); and column 2 contains the results of the non-

linear generalized method of moments IV (GMM-IV) estimation. Comparing the two columns,

the impact of IV estimation is concentrated on rent and the own (0-250 meters) ring stock

coefficients. First, the absolute value of the IV coefficient on rent is much higher than the

ordinary Poisson coefficient. Higher rents besides higher costs (negative effect) reflect better

unobserved contemporaneous and expected future neighborhood amenities (positive effects) in

the ordinary Poisson specification. Also as expected, the IV estimation of the own ring stock

coefficient (.0218) is lower (by 35%) than the ordinary Poisson estimation. A larger number of

19

agencies in the first ring may also reflect higher unobserved and somewhat persistent—from

before 1992—neighborhood amenities affecting both new births and stocks in 1992.

In columns 3 and 4 for 1992-97 births, while the direction of bias in the rent coefficient is

the same, there is no difference in the own ring stock coefficients between IV and ordinary

Poisson estimation. We have three comments on this. First, additional births that occur from

1995 to 1997 are influenced by error drawings for those years which are now separated in time

from the shocks that affected the accumulation of own ring stocks prior to 1992. Second,

discussing bias in terms of absolute coefficients may be misleading. Given a discrete choice

framework, coefficients aren’t absolute profit function coefficients; below we show the degree of

bias in terms of willingness to pay to have more agencies in ring 1 is larger for 1992-97 births

than 1993-94 births: a 3.2 fold increase compared to a 2.4 increase. Third, the IV estimates of the

own ring and rent coefficients are basically the same for 1992-97 births and 1993-94 births. In

general we are going to rely on 1992-97 results because the sample of births is larger; but

throughout results for the two samples are always very similar.

3.2 Interpreting Results

There are two interpretations of the coefficients of the estimated profit function, in the

discrete choice framework. First, under a count model interpretation, coefficients indicate by the

percent by which a change in a covariate raises or lowers the expected number of births in a

tract.14 To help evaluate relevant magnitudes, Table B-1 in Appendix B reports means and

standard deviations of covariates. Second, we can monetize effects on profits utilizing the rent

coefficient. We can calculate for each ring stock magnitude what percentage increase in rent per

square foot of office space a firm would be willing to pay to have one more neighbor, or we can

ask how many standard deviations increase in unit rent a firm would pay for a one standard

deviation increase in neighbors. Later, for a very small sample of firms, we will actually look at

the effects of covariates on ex post absolute profits, to give a sense of how the estimated

coefficients of our discrete choice model compare in absolute magnitude to actual parameters.

Column 4 of Table 4 says that a 1% increase in rents leads to a 2.6% decrease in expected

numbers of births in a tract, indicating births within Manhattan are very sensitive to rents. In

14 Equivalently, in a multinomial logit framework, the effect is to change the probability of a birth by a particular percent (assuming the base probability of a birth in any tract is small)

20

particular, a one standard deviation change (.17) in the rent variable leads to a 44% drop in

expected births in a tract. This magnitude helps explain the shifting pattern of advertising agency

locations in Manhattan in response to relative rent changes across tracts over the last 15 years, as

noted earlier.

On the benefits of having more neighbors, column 4 shows that, as we move from the

first ring out in 250 meter increments, the coefficients are .020, .023, .0042 and then zero beyond

750 meters. Ring 1 and 2 effects are always similar, but at 500 meters, the drop to the third ring

is always large, with the coefficient being small and insignificant in IV estimation in column 4,

although in column 2 it is significant at a 10% level. Any inferred networking effects end with

ring 3. For the hypothesized networking effects, close spatial proximity is critical and thus so is

having greater densities of establishments in commercial neighborhoods. Based on numbers of

advertising agencies per ring in Appendix B and changes in ring areas as we move out, the

number of advertising agencies per unit spatial area does not decline as we move to outer rings,

so there is no sense that rings 4 and 5 have anything less to potentially offer the typical firm in

the sample; they are just too far away. Interactions occur primarily within 500 meters, perhaps a

15-20 minute journey of elevator rides and walking during the day within Manhattan with its

crowded conditions. This finding is robust to specifications of ring distance measures, as

discussed below. And as noted earlier each additional ring encompasses 3-4 more tracts, which is

a small increment.

One question concerns why the ring 1 effect doesn’t dominate the ring 2? First, it could

be that commuting costs are non-linear: people may be indifferent between a 5 minute and 12

minute walk, but not between those and a 20 minute one. Second, although we don’t think there

are shopping-center or market potential externalities, there may still be a poaching problem.

While for a firm, it is great to have a network “member” 5-10 minutes walk down the block,

having an agency right next door could “advertise” a neighbor’s features to current clients who

may choose to visit the agency occasionally.

In terms of magnitudes, if the number of neighbors in ring 1 increases by one, the

expected number of births increases by 2%. For typical variations, the standard deviation of ring

21

1 neighbors is 10 and of ring 2 is 20 (and ring 3 is 49).15 If we increase the number of immediate

neighbors by one standard deviation that raises the expected number of births by 20% and if we

increase the number in ring 2 by a standard deviation that raises the expected number of births in

the own tract by 46%. Even for ring 3, with its point estimate of .004, the effect of a one standard

deviation increase on expected number of births is 20%. These are very large absolute effects.

In terms of willingness to pay higher rent in order to enhance networking possibilities,

column 4 indicates (the number given in square brackets) that a firm would be willing to pay 0.8

percent (.0198/2.57x100) higher rent per square foot to have one more neighbor in the first ring.

That willingness to pay for a larger network is similar for ring 2. For ring 1, firms would be

willing to pay just under a half a standard deviation increase in unit rent (or .078 of .17) for one

standard deviation increase in immediate neighbors. Or if a firm moves from the fringe with no

neighbors to the height of the action in our sample with 50 immediate neighbors, it would be

willing to pay over 2 standard deviations in unit rental increases. Of course, that move could

bring more neighbors in rings 2 and 3, as well. There is a substantial rent-networking trade-off in

choosing locations for advertising agencies within Manhattan.

These networking-scale results are based on a linear formulation of ring counts. We

experimented with a quadratic formulation and a log-linear one. We don’t highlight these results

because their IV versions are statistically weak.16 But for ordinary Poissons both formulations

“work”. Under a log-linearity, the coefficients (and standard errors) for the first three rings are

.391 (.066), .205 (.070), and .140 (.0603); these are very high scale elasticities. Quadratic scale

effects are also large. For example for the first ring, the coefficients on the linear and squared

terms are .0470 and -.000509 (both significant at 10% or better level). If we increase the number

of neighbors by a standard deviation from 6 to 16, in the first ring expected births rise by 41

percent. Positive marginal effects persist until 45 neighbors, which interestingly coincides with

the highest concentration of agencies in Manhattan (i.e., 51 agencies in the first ring).

Finally, the remaining variable is the number of commercial establishments of all types in

a tract. This could represent a local agglomeration economy, or a market potential and shopping-

15 Note the standard deviation and average of ring counts rise as we move out. It is expected since the ring area ( 1 2 1 2( )( )r r r rπ − + ) rises as we move out where 1 2( )r r− is a constant (the 250 m. increment) but 1 2( )r r+ rises with distance. Table B-1 in Appendix B reports the agency count and its standard deviation in our sample for the rings. 16 No coefficients are significant in the quadratic and only the coefficient for the first ring is significant in the log-linear formulation.

22

center effect. We will show below that such effects do not extend beyond the first ring. The

variable could also represent the potential number of slots available for a new agency, where

more slots offer better idiosyncratic matching of individual establishments to the neighborhood

in nested logit context, as discussed above. For the latter, the coefficient implies a significant

positive correlation between error drawings within tracts. Overall, the effect is fairly large: a 1

percent increase in the number of commercial establishments in the tract raises expected births

by 0.5%. More particularly from a mean of 400 establishments per tract, a firm would be willing

to pay .05% increase in price per square foot of office space for one more establishment.

3.3 Identification

The instruments are listed in the footnote of Table 4. These instruments include 1970

housing market variables and counts of advertising agencies for different rings in 1977. All 1992

covariates are treated as endogenous, except for own ring total commercial establishments. By

all criteria total spaces for commercial establishments in a tract are exogenously given by New

York zoning constraints and not sensitive to immediate market conditions. First stage regressions

indicate the instruments are strong, partial and minimum F statistics are reported at the bottom of

the table, and the partial R 2’s average .50.

The chief issue is the validity of using 1977 ring stock variables as instruments,

especially for the first ring. The issue is that we have asserted some persistence in location

patterns of advertising agencies over time due to moving costs which makes 1977 stocks possible

instruments, but that tract unobservables change over time so 1992 residuals are uncorrelated

with 1977 stocks. First, we note that the own ring stock first stage regression is the weakest of

first stage regressions with the lowest partial F (7.92).17 Second, Sargan tests cannot reject

validity of instruments; in fact, overall Sargan values in columns (2) and (4) of the table are very

low, indicating that the specification and our assumption that current errors are orthogonal to

past stocks are valid. Nevertheless, we conducted many other experiments.

As an instructive, informal test, we add 1977 stocks for five rings to the ordinary Poisson

in column 3 of Table 4. The results are in the text table where the first row reports ring

17 The coefficient on the 1977 stock variable in the first stage regression for 1992 own tract ring stock only has a t-statistic of 1.44, s predictability comes from other instruments.

23

coefficients for the ordinary Poisson and the second row those coefficients when 1977 ring stock

instruments are added in. If 1977 stocks are not valid instruments (are correlated with error

‘92 Ring 1

‘92 Ring 2

‘92 Ring 3

‘92 Ring 4

‘92 Ring 5

’77 Ring 1

’77 Ring 2

’77 Ring 3

’77 Ring 4

’77 Ring 5

Base case (Table 4)

.021** (.0043)

.017** (.0026)

.0039** (.0014)

-.0015 (.0016)

-.0022 (.0015)

Instruments added

.023** (.0049)

.017** (.0030)

.0052** (.0018)

.0034 (.0023)

-.0015 (.0019)

.00046 (.0034)

.0035 (.0028)

-.0031*(.0017)

-.0037** (.0018)

-.0025 (.0016)

terms) and 1992 stocks are also correlated with error terms, the coefficients of 1992 stocks

should decline. The table shows for the key variables for the first three rings, no decline occurs.

Moreover, the 1977 stocks have zero coefficients for the first two rings. We also divided the

1992 stocks into new (post 1982 birth date) versus old firms, inserting each as separate sets of

covariates. In ordinary Poissons, the former strongly dominate the latter because the new stocks

are more likely to be correlated with current shocks. In IV estimation, with so many variables the

results are not statistically strong, but the pattern of dominance by new stocks disappears as it

should, because IV estimation removes any correlation of new stocks with current shocks.

In Table 5, we explore the validity of instruments and our approach further. Column 1

repeats the base case from Table 4, column 4. Then we explore the use of an alternative

instrument and the issue of whether births replicate deaths.

Alternative Instrument. In column 2 of Table 5, we replace the 1977 own ring stock instrument

with the own ring count of headquarters in 1977 as discussed earlier. That leads to a weaker first

stage F-statistic (6.1) for the own ring stock; although the Sargan p-value declines, it remains a

satisfactory .41. The rent and relevant ring stock coefficients are both somewhat lower under this

specification with for example the own ring stock coefficient dropping 18%; however the

willingness to pay magnitudes for additional neighbors are similar to the base case (even higher

for the own ring).

Deaths. A concern with using the 1977 own ring stock variable as an instrument is that births

may simply replace deaths and location patterns replicate over time. Here we examine deaths,

estimating an equation for agency deaths from 1992-1997 using the same set of covariates.

Results are reported in columns 3 and 4 of Table 5, where, since this is a new model, we present

both ordinary and IV results. Deaths are agencies that disappear entirely (not movers within

Manhattan). Based on Davis, Haltiwanger, and Schuh (1996), individual deaths occur for

idiosyncratic reasons, in particular special circumstances involving the entrepreneur. As such,

24

agencies’ deaths by tract should be roughly proportional to tract stocks, so the first ring

coefficient will be significant. [We note the coefficient of 0.076 on the first ring stock suggests

that if we go from zero to mean stock per tract of about 6, deaths increase in the tract by 45%.

That means about half of the tract stock dies, which is what happens on average.] However, if

births simply replace deaths so tract patterns are replicated over time, the death equation

estimates should mimic the birth ones. The ordinary Poisson (column 3) and in particular the IV

estimates (column 4) show this is definitely not the case. In column 4, the effects of rent, total

establishment, ring 2, and ring 3 disappear completely, neither mimicking effects for births nor

offering instructive results on deaths. The equation fails the Sargan test, which as a joint test on

instruments and specification, indicates the death model is poorly specified.

3.4. Robustness to inclusion of related activities.

Following the idea of backward and forward linkages, in Table 6 we explore the effects

of proximity in Manhattan to activities other than advertising agencies. Column 1 repeats our

base case estimates from column 4 of Table 4. Then we explore the effect of various other

controls, on own ring stock and rent coefficients.

Control for Total Commercial Activity. Column 2 shows the effect of dropping the control for

total commercial activity in the own tract. Doing so strongly increases the coefficient of and

willingness-to-pay for the first ring stock of SU agencies. This suggests that the control does

capture important aspects of having more commercial activities in the own tract, whether they

reflect some type of local agglomeration economy beyond the own industry, or greater choice

among own tract commercial rental properties in the nested logit approach. Column 3 shows that

having more commercial establishments in rings beyond the first ring, or own tract, generates no

benefits, consistent with an underlying nested logit specification. If total commercial

establishments measure market potential externalities, it would be odd to have such benefits

dissipate entirely within 250 meters.

Related Economic Activities. We explore a variety of more specific aspects of access to other

commercial activities: activities that represent clients (headquarters), inputs (graphic services),

and outlets (media). For these we had difficulty finding significant effects and widened the net

and strengthened any effects by combining the first two rings and defining possible effects for

25

these other activities to occur in 0-500 meters. Of the possibilities, only access to broadcasters

matters.

For clients, we have already discussed in the section on instruments that, by 1992, access

to buyers, or headquarters, is unrelated to location choices within Manhattan. Thus the

coefficient (not reported in Table 6) on the count of headquarters within 500 meters is negative

and insignificant in both ordinary Poisson and IV estimation. For graphic services, in IV

estimation, the partial F in first stage regressions for this variable (after adding 1977 stocks on

headquarters, broadcasters and graphics design to the list of instruments) is a dismal 2.26. Thus

we considered only ordinary Poisson results. As reported in Table B2 of Appendix B, there is no

effect on ordinary Poisson results from including graphic services nearby and it has a coefficient

of zero.

In column 4 of Table 6, we examine access to broadcasters who place advertisements.

We measure broadcasting scale by total employment in broadcasting; this works better than

using a count of establishments given the vast dispersion in broadcast establishment sizes. The

IV results show that a one standard deviation increase in broadcasting employment (1785) raises

expected births by 18%, a noticeable effect (although the ratio of standard deviation to mean is

high (3.8) reflecting zeros in many rings, with very high concentrations in a few). Inclusion of

this variable leaves other coefficients unaffected. In summary, in Table 6, we find no evidence

that inclusion of other suspected relevant local activities diminishes the estimated benefits of

having more advertising agencies in successive rings.

We also looked at establishments of multi-unit advertising agency firms, although these

results are not detailed here. Ordinary Poisson results suggest that having multi-unit advertising

agencies nearby helps in the first ring (significant coefficient of .027) and hurts in the second;

but IV results give small and insignificant coefficients (first ring coefficient of .0045). This fits

with results in the Appendix that show multi-unit agencies do not derive benefits from single unit

firms.

Address Signaling Effects. Finally we looked for address signaling effects. As new firms that

are just starting out, we don’t expect address effects for births, in contrast perhaps to huge

established firms that do business with America’s largest corporations. In the last column of

Table 6 we add a dummy for the five tracts below Central Park that contain Madison Avenue.

The basic IV results are little affected and the coefficient on the Madison Avenue dummy, while

26

positive, is small and insignificant. If we narrow the dummy to the 3 classiest Madison Avenue

tracts the coefficient becomes negative (and insignificant). We also experimented with distance

measures, such as distance to Rockefeller Center; again no effects.

3.5 Robustness to specification

In Table 7, we carry out other robustness checks for our basic specification and sample of

tracts. These include eliminating “irrelevant” tracts as choices, altering ring size definitions,

looking at the determinants of actual profits for a small sample of firms, and considering

alternative measures of own industry information spillovers.

Relevant Tracts. First, in Table 7a, we restrict the sample of tracts to 138 tracts that have ever

had a presence of advertising agencies since 1977. That eliminates 26 tracts that may not be

relevant to advertising agency location decisions. Coefficients in column 1 are similar to those in

Table 4. While absolute values tend to fall (the rent coefficient is about 25% lower and the ring 1

stock variable coefficient is about 11% lower) as expected18, the willingness to pay for neighbors

rises modestly.

Ring Specification. In column 2 of Table 7a, we turn to our specification of ring distances. We

defined rings by equal increments of distance to reflect equal changes in contact costs. Thus, the

contact (commuting) costs to an agency in the second ring are twice those for the first ring. We

believe this in the best way to specify distance decay. However, one could define rings to have

equal areas, for example, rather than equal incremental distances from the own tract. In column

2, we define equal area rings (see the bold numbers in the column) of 0-350, 350-500, 500-610

and 610-707 meters. Results do differ somewhat from the equal distance specifications. First, the

ring effect in moving from ring 1 (ending at 350 meters) to ring 2 increases significantly. This

suggests the poaching issue may be stronger than hypothesized above. While the scale effects

persist for 500-610 meters, they disappear at 610 meters, shorter than our 750 meters before.

Actual Profits. In columns 3-5 of Table 7a, we consider the absolute magnitudes of our

estimated, normalized coefficients. For the 30 single unit advertising agencies in Manhattan for

which 1992 Business Expenditure Survey data exist, we calculate profits as sales minus all

operating expenses including rent. We regress the log of profits on the same covariates as our

18 In general, the coefficients in discrete choice models— the generalized exponential model in our case—are scaled by the standard deviation of the residuals. As we excluded the tracts with no births, we expect that the standard deviation of the residuals and therefore the magnitude of coefficients decrease.

27

basic births model in order to estimate non-normalized parameters of the profit function directly.

Given the small sample size, we focus on OLS results. While under a nested logit specification to

the location-count problem, we should control for the total number of establishments in the tract,

that argument is no longer relevant per se for this direct estimation of profits. However, the

number of establishments does control for possible general local agglomeration economies, as

noted earlier. In columns 3 and 4 of Table 7a, we show OLS results with and without the control;

the coefficient on this control is weak (relatively speaking in a context where all results are weak

with the small sample size). In these columns, the pattern of coefficients is similar to those in our

basic birth model, with a large negative rent coefficient and high benefits of near neighbors,

which decline sharply with distance. The results hint at what absolute coefficients might be,

compared to our count estimations. Coefficients for the first two ring stock variables are at least

twice ordinary Poisson coefficients. The last column shows 2SLS results where we instrument

for rents and the first 2 ring stock variables, without the control on total establishments.

Alternative Spillover Measures. The final issue is that we have measured advertising agency

networking and information spillover effects based on just a count of enterprises. The idea is that

each enterprise represents a source of information spillovers and a networking opportunity. It

could be that employment counts are a better representation of networking opportunities, but

how can we distinguish the two? One way is to decompose total SU ring employment into a

count of establishments and average employment size to see if they yield similar effects. Such a

decomposition serves double duty, allowing for the possibility that, even if establishment counts

accurately record the opportunities for networking, quality of neighborhood establishments may

matter. If size is a measure of quality as discussed in section 2.1, this decomposition allows us to

test for the existence of either employment sources of externalities or quality of neighbor effects.

The basic problem in proceeding with decomposition is that for IV estimation, while we

can still instrument for establishment counts, we have no strong instruments for average

establishment size. Adding to our instrument list 1977 variables on average agency size by ring

produces first stage regressions for average establishment size with F-statistics under 2.0.

Corresponding if we change the spillover covariates to total advertising agency employment

counts in 1992, in first stage regressions all variation in total employment counts is explained

just by 1977 establishment count variation. We report on two sets of results.

28

First in Table 7b, column 1 for IV estimation we replace establishment counts with

employment counts of advertising agencies (log employment) in rings. Given our comments on

first stage regressions in the previous paragraph, the partial F-statistic for the own ring

employment stock regression is only 3.7. The results show positive employment effects for the

first three rings (now in elasticity form: percent changes in births for a percent change in

employment stocks).

Second, we report on two decomposition attempts for total employment, for ordinary

Poissons only, since we have no instruments for establishment size measures. While the basic

decomposition is into the count of agencies and average employment per agency to distinguish

the effects of more firms versus either more employees and/or larger and higher quality firms,

averages can be a noisy measure of quality driven by outliers. Thus we also experimented with

median firm employment by ring, as the quality-establishment size measure. For the ordinary

Poisson estimation, we define just three rings. The last column defines the base where we control

for ring establishment counts, without the employment measures. Adding median employment

by ring in column 2 of Table 7b shows no effect on establishment count coefficients relative to

the last column and yields insignificant coefficients for the employment size measures. However,

using average size measures does reduce the establishment count coefficients for rings 1 and 2

by 18 and 30% in column 3 compared to the last column (with little effect on the rent

coefficient); and average employment size variables are significant for ring 1 and almost so for

ring 2. Thus there is at least a hint that employment sizes as well as establishment counts matter.

We can’t sort this out for IV estimation; and, conceptually, we prefer the establishment count

formulation.

3.5 Spatial Correlation.

When using data at this fine level of detail, and in addition when asserting within tract

correlation of error terms, the issue of spatial correlation of error terms across tracts arises.

Before addressing that directly, note that in Table 4, when we instrument for ring variables, we

instrument for them all. So if the shock to the own ring affects ring stocks and shocks are

correlated across tracts, then the shock to the own ring could be correlated with ring 2 or ring 3

covariates. Instrumenting with historical variables takes care of that problem; even if the

29

instruments are historically spatially correlated, the concern is that the shocks today are

independent of these historical variables.

The key issue with spatial correlation is the potential for biased estimates of standards

errors. Thus we checked for spatial correlation, using a Moran I test on the residuals from both

the ordinary Poisson and the GMM models in Table 4, columns (3) and (4). The Moran I test

examines whether a regression of neighbors’ average residuals on own tract residuals is

significant. Near neighbors in the averaging are given a weight 1 and others a weight 0. In

denoting what is a near neighbor, the usual procedure is to think of each tract as having 8

neighbors on a lattice. We use a radius from the own tract that gives an average of 8.3 neighbor

tracts as neighbors. For this radius, the hypothesis of no correlation cannot be rejected at a p-

value of .10 for an ordinary Poisson. The result for GMM coefficients is similar—no rejection

with a p-value of .11.19 Thus within our data, while, given tracts are neighborhoods, there may

be within tract correlation, the problem of cross tract correlation does not seem compelling.

4. Conclusions and Extensions The fundamental findings in the paper are (1) at a micro spatial level, scale externalities

are very large but they dissipate very quickly with distance for advertisers and are gone by 750

meters and (2) benefits of being in a better neighborhood are capitalized into land rents. The

effects strongly suggest networking, or information “spillover” effects. The results raise two

issues. First use of wage equations to estimate urban agglomeration economies ignores the fact

that some portion of benefits are capitalized into both the overall level of commercial rents in a

city and rent variation across space within commercial sectors. Second the finding of quick

spatial decay of information spillover benefits for advertising agencies begs the question of what

one measures with more aggregate data, such as MSA scale effects. Is it labor market

externalities (operating at the level of the entire labor market) or is it that having more

advertising agencies in a city allows a greater choice of clusters within a city?

The idea that firms within a city in choosing locations are trading-off rent costs versus the

benefits of being “at the center of action” in a large cluster versus operating on the fringes raises

19 Only if we drop the radius so the average number of neighbors is 3.3, can the hypothesis of no correlation be rejected (p-value of .015). We also regressed residuals of pair-wise neighbors on each other, getting exactly the same pattern of results. For those we also looked at a sample requiring all pairs to both have counts greater than 1 (worrying about the discreteness problem with count data). The same results apply.

30

the issue of heterogeneity. Which firms are choosing to pay higher rents in order to have better

neighborhood networking – are these higher or lower quality firms? And as we noted earlier,

heterogeneity raises the possibility of a separating equilibrium where high quality firms segment

themselves away from low quality firms by locating in tracts that are too expensive for low

quality firms. That notion may in part explain another puzzle: if location choices of advertising

agencies are driven by the desire to network just with other agencies, why not cluster in low rent

areas on the fringes of Manhattan, instead of clustering in higher rent areas? Partly as Tables 4

and 6 revealed advertising agencies do value being in tracts with high concentrations of overall

commercial activity and do value being near broadcasters. But the notion of separating equilibria

adds an additional possibility. In an earlier version of the paper, we looked at heterogeneity in

detail, with the limited sample available. Results are suggestive results, but there is limited

precision, so our presentation here is brief.

In introducing observed heterogeneity, we assume from the model that better quality

firms are larger, so we can treat size as an index of quality. We can’t introduce heterogeneity for

births, because we have no ex ante information on their quality (and the firms themselves may

not have learned their quality); and ex post size measures are likely to be influenced by

contemporaneous error drawings. However we have a sample of 95 movers who change location

between 1992 and 1997 within Manhattan. These experienced firms have informed notions of

their own quality; and, for 82 of them, we have a size measure of quality-- payroll in their prior

location in 1992. That variable should be uncorrelated with the error drawings in the new

location. We note payroll and employment information from social security records are the only

non-imputed census variables on firm economic magnitudes available for more than a handful of

firms. We use payroll (where there is a uniform wage for uniform workers in Manhattan), to

allow better quality firms to also hire better quality, or higher paid employees, but there is little

difference in results.

To incorporate firm heterogeneity into the empirical estimation, we follow the approach

in Berry (1994) and Bayer, McMillan, and Rueben (2004) and postulate a heterogeneous

coefficients model. In the profit function appearing in (6) we assume the coefficient vector φ for

neighborhood covariates now takes the form

0 1k kzφ φ φ= + . (9)

31

kφ are firm k ’s specific coefficients on the jX covariates, where those coefficients vary with

firm characteristics, kz , which in this case will be a measure of firm quality. For scale effects in

equation (6), we are allowing ln [ ]E ⋅ to change as eitherγ or the expected u drawing from

equation (5) vary with firm quality; for rental effects we are allowing α to vary. Once we

introduce quality issues, in our framework we should also ask whether firms value quality of

neighbors, as well as counts. For an incoming firm, payroll levels of existing firms are a signal as

to neighborhood quality. As such, we use the median firm payroll by ring in 1992 as a measure

of neighborhood quality, recognizing this is an approximation.20 As we already know, however,

for ordinary Poisson’s only the first ring has significant median neighborhood size effects and in

IV work we have weak instruments. In what we discuss, we only control for ring one quality of

neighbors.

Given neighborhood characteristics are endogenous and we have firm specific

characteristics, we employ a two stage identification strategy. First, we estimate (7) with tract

fixed effects, so the basic logit equation for firm k considering location j is

1

1

exp[ ]exp[ ]

kj jk

j ki i

i

d z XP

d z Xφφ

+=

+∑ (10)

In (10), jd is a neighborhood dummy variable, which controls for all neighborhood

characteristics, observed or not. As long as the kz are exogenous, the 1φ are then identified. But

that leaves the base coefficients, 0φ , for the jX covariates. These are given by the equation

0j j jd b Xφ= + , (11)

where the LHS dummy variables are estimated in (10). To identify the 0φ , given the jX are

correlated with unobserved neighborhoods contemporaneous attributes, equation (11) is

estimated by 2SLS, using as instruments the variables we used before in estimating the moment

condition (8).

20 The issue is characterizing quality by a single measure of size. For example, this is a correct measure if the stock of firms is spread across neighborhoods so as to perfectly segment the line representing the quality types in the market. With perfect segmentation, the lowest quality firm in a neighborhood is at least the same quality as the highest quality firm in the next best quality neighborhood. Then greater size signals a first degree stochastic dominant improvement in the distribution.

32

In estimation, because of the limited sample, to sharpen the reported results, we truncate

ring effects at 3 rings. Results are given in Table 8. The first two columns give ordinary logit

results for movers. Column (1) presents results without heterogeneity, where scale is for 3 rings

and quality for 1. Column (2) presents the ordinary conditional logit results for the full model in

(7), amended by (9) to allow for heterogeneous coefficients, so all variables in column (1) are

now additionally interacted with the quality measure— payroll size in the prior location. If we

evaluate these at median size of movers, results in column (2) are similar to those in column (1)

on all variables. For that comparison,21 we use a “synthetic” median, which is 4.75; this is the log

of the mean size of the 10 firms around the median. However there are enormous differences by

firm quality. In particular ordinary logit results suggest better quality firms are much less

sensitive to rent differentials and much more sensitive to immediate ring 1 neighbors. We

explore this next in columns (3) and (4), where issues of endogeneity are handled.

In column (3) we present results of the tract fixed effect logit model in equation (10).

Notice the coefficients for these interactive variables are similar to those in column (2), but with

stronger rent and ring 1 scale coefficients. In column (4), we recover the 0φ coefficients by

estimating equation (11) by 2SLS. For this, the sample size is only 40 tracts which raises a small

sample problem, so in column (5) we also give the OLS results. To evaluate effects of any

variable, we must combine the results in columns (3) and (4).

Results indicate that higher quality firms are less concerned about rent costs. The strongly

negative rent coefficient for a small firm declines in absolute magnitude to -.36 for the median

size mover. This could indicate that there is also a fixed cost portion to rent which effect

diminishes with firm size. Second, higher quality firms benefit more from scale externalities, at

least in ring 1. The ring 1 scale effect becomes positive at a very small firm and then rises

sharply, so that at median size it is a high .034 (cf. Table 4 coefficient of .02) and at median size

plus one standard deviation of size it is a huge .078. High quality firms are much more likely to

choose neighborhoods with more agencies. Given the rent effect declines with firm quality, the

willingness to pay in terms of rents for more neighbors is very high for higher quality firms.

Point estimates suggest the median firm is willing to pay a 9.4% increase in unit rent to have one

more neighbor, but with our limited sample size we lack precision. Finally, in the Table, results 21At median size, the establishment, rent, and ring 1-3 scale and ring quality (heterogeneous) coefficients are respectively .21, -.36, .022, .015 and .0075, while those in column (1) are .30, -.093, .020, .018, and .0059, indicating here a formulation without heterogeneity tends to capture the coefficients of the “representative” firm.

33

in all columns on neighborhood quality for our movers, as measured by the median payroll size

of existing establishments in 1992 in the first ring, suggest quality doesn’t matter. Of course we

are looking within New York, which nationally has a high end cut of larger and presumably

higher quality firms (Arzaghi, 2005).

34

References

Arzaghi, Mohammad (2004), “A Model of Knowledge Sharing and Network Formation Among Advertising Agencies,” Working Paper, Brown University.

Arzaghi, Mohammad (2005), “Quality Sorting and Networking: Evidence from the Advertising Agency Industry,” Working Paper, Brown University.

Bala, Venkatesh; Goyal, Sanjeev (2000), “A Noncooperative Model of Network Formation,” Econometrica, 68, 1181-1229.

Bayer, Patrick; McMillan, Robert; Rueben, Kim (2004), "Residential Segregation in Equilibrium," Yale University mimeo.

Berndt, Ernest R; Alvin J. Silk (1994), “Costs, Institutional Mobility Barriers, and Market Structure: Advertising Firms as Multiproduct Firms,” Journal of Economics and Management Strategy, 3, 437-450.

Berry, Steven (1994), "Estimating Discrete-Choice Models of Product Differentiation," Rand Journal of Economics, 25, 242-262.

Davis, S., J. Haltiwanger, and S. Schuh (1996), Job Creation and Destruction, Cambridge, MIT Press.

Dudey, Marc (1990), “Competition by Choice: The Effect of Consumer Search on Firm Location Decisions,” American Economic Review, 80(5), 1092-1104.

Duranton, Gilles; Overman, Henry (2005), "Testing for Localization Using Micro Geographic Data," Review of Economic Studies, 73, 1077-1106.

Ellison, Gllen; Glaeser, Edward (1997), "The Geographic Concentration of Industry: A Dartboard Approach," Journal of Political Economy, 105, 889-927.

Glaeser, Edward and David Mare (2001), “Cities and Skills,” Journal of Labor Economics, 19, 316-342

Glaeser, Edward; Gyourko, Joseph (2002), "The Impact of Zoning on Housing Affordability," Harvard Institute of Economic Research Paper # 1948.

Goyal, Sanjeev; Moraga-Gonzalez, Jose Luis (2001), “R&D networks,” RAND Journal of Economics, 32(4), p 686-707.

Guimarães, Paulo; Figueirdo, Octávio; Woodward, Douglas (2000), "A Tractable Approach to the Firm Location Decision", Review of Economics and Statistics, 85, 201-204.

Hameroff, Eugene (1998), The Advertising Agency Business, New York: McGraw-Hill.

Henderson, J. Vernon (2003), "Marshall's Scale Externalities," Journal of Urban Economics, 53(1), 1-28.

Holmes, Thomas (2005), “The Location of Sales Offices and the Attraction of Cities,” Journal of Political Economy.

Jackson, Matthew; Wolinsky, Asher (1996), “A Strategic Model of Social and Economic Networks,” Journal of Economic Theory, 71, 44-74.

35

Jaffe, Adam B.; Trajtenberg, Manuel; Henderson, Rebecca (1993), "Geographic Localization of Knowledge Spillovers as Evidenced by Patent Citations", Quarterly Journal of Economics, 108, 577-598.

Lucas, Robert E. Jr. and Estaban Rossi-Hansberg (2002), “On the Internal Structure of Cities,” Econometrica, 70, 1445-1476.

McMillen, Daniel P. and Larry D. Singell, Jr. (1992), “Work Location, Residence Location, and The Intraurban Wage Gradient,” Journal of Urban Economics, 32, 195-213.

Mullahy, John (1997), "Instrumental Variables Estimation of Count Data Models, With an Application to Models of Cigarette Smoking Behavior," Review of Economics and Statistics, 79(4), 586-593.

Rosenthal, Stuart; Strange, William (2004), "Evidence on the Nature and Sources of Agglomeration Economies,” in Handbook of Regional and Urban Economics, J.V. Henderson and J-F Thisse (eds.), North Holland.

Rosenthal, Stuart; Strange, William (2003), "Geography, Industrial Organization and Agglomeration," Review of Economics and Statistics, 85, 178-188.

Schemetter, Bob (2003), Leap! A Revolution in Creative Business Strategy, Hoboken, NJ: John Wiley & Sons, Inc.

Windmeijer, Frank; Santos Silva, Joya (1997), "Endogeneity in Court Data Models: An Application to Health Care," Journal of Applied Econometrics, 12, 281-294.

36

Table 1. 1992 Economic Census Numbers on Concentration in Advertising*

Advertising agencies All other business services Share of total No. of units Share of total No. of units Counts of establishments with under 5 employees 58% 6890 52% 133,380

Sales by establishments with under 5 employees 14% 6890 10% 133,380

Counts of establishments with 100+ employees 1.3% 156 4.1% 10,590

Sales by establishments with 100+ employees 33% 156 40% 10,590

Counts of single-unit firms 90% 12,453 81% 236,212

Sales by Single-unit firms 55% 12,453 45% 236,212

* The establishment numbers cover just “operating establishment”. The firm numbers cover all establishments. Numbers are from the 1992 Census of Services.

37

Table 2. New York City’s Role in the Advertising agency Industry 1992

Share of US establishments 7.3% Share of US employment 20% Share of US receipts 24% Share of media billing 31%

Table 3. New York City Births

1997 Stock Born Before ’92-97 Births ‘73 ‘88 single unit est. 1088 104 357 545

38

Table 4. Birth Models.

Birth ’93-’94 Poisson

Birth ’93-’94 IV-GMM

Birth ’92-’97 Poisson

Birth ’92-’97 IV-GMM

(1) (2) (3) (4) Ln(total no. of .411** .530** .462** .498** establishments) (.137) (.216) (.0756) (.117) Ln(rent/sq. ft.) -1.72** -2.74* -.819** -2.57** (.572) (1.46) (.289) (.773) Stock of agencies .0334** .0209* .0206** .0198** 0-250m. (.00817) (.0119) (.00430) (.00758) [Will-to-pay/sq. ft. (%)] [1.9] [.76] [2.5] [.77] Stock of agencies .0147** .0191** .0167** .0228** 250-500m. (.00494) (.00953) (.00264) (.00538) Stock of agencies .00451* .00648 .00387** .00419 500-750m. (.00250) (.00500) (.00137) (.00259) Stock of agencies .000702 -.00336 -.00149 -.00149 750-1000m. (.00281) (.00639) (.00158) (.00274) Stock of agencies -.00577** -.00519 -.00219 -.00241 1000-1250m. (.00272) (.00605) (.00147) (.00251) N 164 164 164 164 Pseudo 2R .343 — .504 — Sargan [p-value] — 3.88 [.694] — 1.71 [.945] 1st stage avg. 2R — .50 — .50 avg. F; (min F) — 13.0; (7.9) — 13.0; (7.9)

The instrument list is distance to Rockefeller Center, 1970 log of total housing units, 1970 share of housing units in buildings with less than 5 units, stocks of single unit advertising agencies in each of the 5 rings in 1977, total count of all establishments in each of the first four rings in 1977, and 1992 total number of establishments in the own tract.

39

Table 5. Alternative Instruments and Births versus Deaths

Births ’92-’97

IV-GMM Births ’92-’97

CA Instrument for the First Ring

IV-GMM

Deaths ’92-’97 Poisson

Deaths ’92-’97 IV-GMM

(1) (2) (3) (4) Ln(total no. of .498** .530** .366** .0600 establishments) (.117) (.110) (.101) (.209) Ln(rent/sq. ft.) -2.57** -1.98** .164 -.0121 (.773) (.578) (.364) (.819) Stock of agencies .0198** .0167** .0598** .0756** 0-250m. (.00758) (.00626) (.00578) (.0140) [Will-to-pay/sq. ft. %] [.77] [.84] Stock of agencies .0228** .0169** .00822** .00864 250-500m. (.00538) (.00304) (.00335) (.00617) Stock of agencies .00419 .00401** -.000208 -.00182 500-750m. (.00259) (.00147) (.00176) (.00273) Stock of agencies -.00149 -.00212 .00447** .0111** 750-1000m. (.00274) (.00181) (.00188) (.00352) Stock of agencies -.00241 -.000949 -.00373** -.00182 1000-1250m. (.00251) (.00167) (.00183) (.00199) N 164 164 164 164 Pseudo 2R — — .549 — Sargan [p-value] 1.71 [.945] 6.11 [.411] — 13.0 [.042] 1st stage avg. 2R .50 .36 — .50 avg. F; (min F) 13.0; (7.9) 11.3; (6.1) — 13.0; (7.9)

Except for column 2, the instrument list is distance to Rockefeller Center, 1970 log of total housing units, 1970 share of housing units in buildings with less than 5 units, stocks of SU advertising agencies in each of the 5 rings in 1977, total count of all establishments in each of the first four rings in 1977, and 1992 total number of establishments in the own tract. In column 2, the office rents and first ring counts are instrumented using the distance to Rockefeller Center, two 1970 housing variables, and 1977 first ring headquarter counts. The other rings counts are assumed exogenous.

40

Table 6. Other Covariates

Base Case IV-GMM

No Control for Total Estab.

IV-GMM

Total Estab. in Other Rings

IV-GMM

Broadcasting Employ.

IV-GMM

Madison Avenue IV-GMM

(1) (2) (3) (4) (5) Ln(total no. of .498** — — .563** .498** Establishments) (.117) (.120) (.126) Ln(rent/sq. ft.) -2.57** -3.60** -2.34** -2.60** -2.74** (.773) (1.07) (1.22) (.850) (1.00) Stock of agencies .0198** .0321** .0214** .0221** .0188** 0-250m. (.00758) (.00943) (1.22) (.00916) (.00898) [Will-to-pay/sq. ft. %] [.77] [.99] [.91] [.85] [.69] Stock of agencies .0228** .0341** .0213** .0190** .0236** 250-500m. (.00538) (.00631) (.00738) (.00530) (.00718) Stock of agencies .00419 .00874** .000769 .00281 .00426 500-750m. (.00259) (.00322) (.00604) (.00261) (.00288) Stock of agencies -.00149 .00450 — -.00298 -.00122 750-1000m. (.00274) (.00416) (.00292) (.00315) Stock of agencies -.00241 -.00484 — -.000954 -.00248 1000-1250m (.00251) (.00331) (.00290) (.00276) Ln (total est) .379** 0-250m (.178) Ln (total est.) -.0529 250-500m (.0845) Ln (total est.) .251 500-750m (.370) Broadcast employ. .000103** 0-500m (.000035) Madison Ave. dummy .0558 (.339) N 164 164 164 164 164 Sargan [p-value] 1.71 [.945] 6.03 [.537] 3.29 [.655] 1.57 [.954] 1.36 [.968] 1st stage avg 2R .50 .60 .33 .55 .44 avg. F; (min F) 13.0; (7.9) 20.0; (13.2) 8.3; (6.1) 16.3; (7.4) 10.3; (5.9)

For column 3, the instrument list is distance to Rockefeller Center, 1970 log of total housing units, 1970 share of housing units in buildings with less than 5 units, and stocks of SU advertising agencies in each of the first three rings in 1977. The log of total count of all establishments in each of the first three rings in 1992 are considered to be exogenous.

41

Table 7a. Robustness: Sample, Ring Definition, actual profits

“relevant”

tracts IV-GMM

Equal area rings

IV-GMM

Profits ‘92 OLS Profits ’92

OLS Profits ’92

2SLS

(1) (2) (3) (4) (5) Ln(total no. of .458** .569** .464 — — Establishments) (.110) (.0934) (1.01) Ln(rent/sq. ft.) -1.96** -1.90** -2.94 -2.30 -2.80 (.719) (.823) (2.78) (2.37) (3.03) 0-350m Stock of agencies .0176** .0129** .0743 .0905** .1088* 0-250m. (.00747) (.00521) (.0561) (.0433) (.0580) [Will-to-pay/sq. ft. %] [.90] [.68] 350-500m Stock of agencies .0178** .0209** .0331 .0322 .0836** 250-500m. (.00524) (.00510) (.0255) (.0250) (.0354) 500-610m Stock of agencies .00399* .0107** -.0248** -.0249** -.0325** 500-750m. (.00224) (.00496) (.0110 (.0108) (.0133) 610-707m Stock of agencies -.00270 -.00159 — — 750-1000m. (.00254) (.00465) Stock of agencies -.00146 — — 1000-1250m. (.00224) N 138 164 30 30 30 Adjusted/Centered 2R — — .122 .150 .132 Sargan [p-value] 1.84 [.934] 5.29 [.508] — — 1st stage avg. 2R .48 .50 — .55 avg. F; (min F) 10.1; (5.4) 14.3; (9.6) — 4.5; (3.5) The instrument list for column 1 is distance to Rockefeller Center, 1970 log of total housing units, 1970 share of housing units in buildings with less than 5 units, stocks of single unit advertising agencies in each of the five rings in 1977, total count of all establishments in each of the first four rings in 1977, and 1992 total number of establishments in the own tract. For column 4, ring sizes are adjusted in calculating instruments; we also only use stocks of agencies for four rings.

42

Table 7b. Robustness: Scale Measures

Births 92-97 IV-GMM

Births 92-97 Median emp.

Poisson

Births 92-97 Average emp.

Poisson

Births 92-97 Poisson

(1) (2) (3) (4) Ln(total no. of .248** .378** .339** .427** Establishments) (.0987) (.0746) (.0740) (.0700) Ln(rent/sq. ft.) -1.07** -.801** -.912** -.887** (.432) (.288) (.280) (.283) Ln(tot. agency emp.) .238** Ln(agency emp.) .0659 .153** 0-250m. (.104) 0-250m. (.00423) (.0570) Ln(tot. agency emp.) .352* Ln(agency emp.) .0292 .112* 250-500m. (.215) 250-500m. (.0881) (.0632) Ln(tot. agency emp.) .190* Ln(agency emp.) .148 .0895 500-750m. (.107) 500-750m. (.109) (.0713) Ln(tot. agency emp.) -.291** — — — 750-1000m. (.118) Ln(tot. agency emp.) .0945 — — — 1000-1250m. (.144) Stock agencies .0215** .0178** .0214** 0-250m. (.00423) (.00452) (.00418) Stock agencies .0142** .00994** .0149** 250-500m. (.00250) (.00283) (.00247) Stock agencies .00261** .00380** .00265** 500-750m. (.00112) (.00121) (.00112) N 164 164 164 164 Pseudo 2R — .50 .51 .500 Sargan [p-value] 2.39 [.793] — — 1st stage avg. 2R .39 — — avg. F; (min F) 8.4; (3. 7)

The instruments for column 1 are distance to Rockefeller Center, 1970 log of total housing units, 1970 share of housing units in buildings with less than 5 units, log of total single unit advertising agencies employment in each of the five rings in 1977, total count of all establishments in each of the first three rings in 1977, and log of 1992 total number of establishments in the own tract.

43

Table 8. Heterogeneity: 92-97 Movers

Logit Logit Logit with tract fixed

effects

Tract Fixed Effects 2SLS

Tract Fixed Effects OLS

Ln(total no. of .296 .617 .867** .872** establishments) (.196) (.496) (.303) (.160) Ln(rent/sq. ft.) -.0933 -3.91** -5.35** -5.31** (.633) (1.82) (1.04) (.482) Stock agencies .0198** -.0449* -.0490** -.0677** 0-250m. (.00929) (.0256) (.0190) (.00827) Stock agencies .0181** .0166 -.00297 -.000387 250-500m. (.00575) (.0161) (.0142) (.00443) Stock agencies .00592** .0193** .0139** .0141** 500-750m. (.00253) (.00677) (.00495) (.00191) Ln(quality) .259** .336 -.320 .125 0-250m. (.0970) (.262) (.271) (.0841) Ln(size)*ln(est) -.0859 -.193 (.0895) (.122) Ln(size)*ln(rent) .748** 1.05** (.313) (.400) Ln(size)*ring 1 .0141** .0175** stock (.00475) (.00578) Ln(size)*ring 2 -.000430 -.000481 stock (.00287) (.00137) Ln(size)*ring 3 -.00247** -.00220* stock (.00123) (.00131) Ln(size)*ring 1 -.0104 -.0176 quality (.0478) (.0723) N [movers] 15580 [95] 13448 [82] 13448 [82] 40 40

2R .209 .0774 .828 .826 Sargan [p-value] 1.10 (.89) For the second to last column, instruments are distance to Rockefeller Center, 1970 ln(total housing units), share of 1970 units in buildings with less than 5 units, 1977 advertising agency stock for rings 1-3, 1977 total establishments for rings 1-2, ln(median advertising agency payroll size in 1977) for first ring, and 1992 total own tract establishment count.

44

Figure 1: Locations of Advertising Agencies in Manhattan.

45

Appendix A. Office Rent Interpolation for NY City

We use the rent for class A office space as an appropriate rent measure for advertising

agencies. The data is provided to us by Torto Wheaton Research. It covers 51 ZIP codes in

Manhattan in 1992. The data includes the asked price for class A office space in dollar per

squared feet. However the data do not provide us with the rent at Census tract level. Thus we

interpolate the rent data for tracts in the Southern Manhattan. First, we assign the rent to the

centroids of the ZIP codes. Second, we constructed a smoothed rent surface using the spline

method (spline estimates values using a mathematical function that minimizes overall surface

curvature, resulting in a smooth surface that passes exactly through the input points). Finally we

interpolated the rent for a tract by averaging the value of the rent surface over the tract.

Figure A-1. Class A Office Rents for Lower Manhattan ZIP codes.

46

Figure A-2. Class A Office Rent Surface for Lower Manhattan.

47

Appendix B. Basic Statistics for Covariates and Other Results

Table B-1. Means and Standard Deviations of Tract Attributes

Tract Attribute Raw Mean Raw S.D. Weighted Mean Weighted S.D. Stock of SU agencies 0-250m. 5.80 9.74 17.5 15.8 Stock 250-500m. 17.3 20.3 36.8 22.2 Stock 500-750m. 37.4 49.0 91.4 69.0 Stock 750-1000m. 34.3 34.6 52.6 33.4 Stock 1000-1250m 54.3 52.9 96.3 61.4 Ln(total no. establishments) 6.07 1.3 7.1 .88 Ln(rent/sq. ft.) 3.32 .17 3.36 .19 Ln(median payroll of SU agencies)

3.30 2.73 5.24 1.98

N 164 164 502 502 Attributes are weighted using the number of births in tracts. Table B2. Graphic Design

Graphic DesignPoisson

Ln(total no. of .463** Establishments) (.0756) Ln(rent/sq. ft.) -.701 (.331) Stock single unit agencies .0191** 0-250m. (.00476) [Will-to-pay/sq. ft. %] [2.7] Stock agencies .0151** 250-500m. (.00339) Stock agencies .00358** 500-750m. (.00142) Stock agencies -.00162 750-1000m. (.00160) Stock agencies -.00192 1000-1250m (.00151) Graphic services .00152 0-500m (.00207) N 164 Pseudo 2R .505

48

Table B-3. Results for births of establishments of multi-unit firms

Ordinary Poisson

Ordinary Poisson

IV-GMM

Ln(total no. of .528** .576** .679 establishments) (.230) (.0767) (.574)

Ln(rent/sq. ft.) -.1.61** -2.25* -2.68

(.807) (1.28) (2.98)

Stock MU’s < .262** .339** .343** 250 m. (.0439) (.0657) (.111)

Stock MU’s .118** .166** .0865 250-500 m. (.0248) (.0416) (.0611)

Stock MU’s .0216 .0307 -.00216 500-750 m. (.0190) (.0197) (.0498)

Stock MU’s -.0461* -.0465* .00271 750-1000 m. (.0257) (.0260) (.0696)

Stock MU’s -.0215 -.0198 -.00187

1000-1250 m. (.0197) (.0211) (.0406)

Stock SU’s -.00219 -.0148 <500 m. (.00147) (.00988)

Broadcast -.0000547

Emp.<500 m. (.0000567)

Constant .0228 1.97 2.00 (3.05) (4.08) (12.0)

N [births] 164 164 164

Pseudo Rsq .460 .468 Sargan p-value .206

Column 1 contains the ordinary Poisson and column 3 the IV for a baseline specification. There

are strong scale effects in rings 1 and 2 for ordinary Poissons and in ring 1 for IV results.

Column 2 shows MU’s do not benefit from being near SU’s. In fact in all formulations we tried

such as having 3 or more SU ring variables in the ordinary Poisson or results under IV

estimation, we never got positive, even modestly significant results for SU variables.